Background
FOLFIRINOX and gemcitabine/abraxane chemotherapy regimens are first-line treatments for
pancreatic ductal adenocarcinoma (PDAC). While FOLFIRINOX generally has superior efficacy,
it is associated with severe side effects that often make treatment intolerable. Moffitt molecular
subtypes of PDAC using RNA expression data have identified a basal-like subtype that is less
responsive to FOLFIRINOX than a classical subtype. The basal-like subtype therefore could be
a better candidate for gemcitabine/abraxane. Here, we developed a proof-of-concept predictor
of basal-like subtype from H&E whole slide images (WSIs) that could be used to rapidly
prioritize cases for further RNA profiling.

Design
WSIs and whole transcriptome RNA-seq data were collected from pancreatic cancer
specimens (N=3,331, Table 1). Basal-like or classical labels were assigned by applying the
Purity Independent Subtyping of Tumors (PurIST) algorithm to RNA-seq input data. An
attention-based convolutional neural network was trained to predict the subtype label from
each WSI in the training set (60% of data). Hyperparameters were selected using the
optimization set (20%) of data. Performance is reported on a holdout set (20% of data). The
entire process of training, optimization, and evaluation was repeated using 5-fold
cross-validation.

Results
The WSI-based deep learning model achieved a mean receiver operating characteristic area
under curve (ROC-AUC) of 0.82 (95% CI 0.77 – 0.86, Figure 1). Performance was similar for
WSIs from within the pancreas (ROC-AUC 0.80, 95% 0.71 – 0.89) and outside of the pancreas (ROC-AUC 0.82, 95% CI 0.77 – 0.88, Figure 2), as well as for large specimens (e.g.
surgical resections; ROC-AUC 0.82, 95% CI 0.72 – 0.91) and small specimens (e.g. core
needle biopsies; ROC-AUC 0.81, 95% CI 0.75 – 0.88), and for both Leica GT450 (ROC-AUC
0.82, 95% CI 0.75 – 0.89) and Philips UFS scanner types (ROC-AUC 0.81, 95% CI 0.75 –
0.87).

Conclusion
We found that molecular subtypes of PDAC are associated with features that can be captured
from H&E WSIs using a deep learning model. Identification of patients likely to have basal-like
PDAC could be used to rapidly identify patients less likely to benefit from FOLFIRINOX therapy
and prompt follow-up RNA expression testing for subtype confirmation. Future work includes
expansion of the study to the institution’s entire set of data and demonstration of the clinical
validity of this model using FOLFIRINOX therapy response data.

Table 1. Available characteristics of the cohort, split according to subtype based on labels assigned via the PurIST algorithm. A chi-square test was performed for each characteristic to determine association with the target label.

VIEW THE PUBLICATION

VIEW THE POSTER